[ARM] Mark MVE_VMOV_to_lane_32 as isInsertSubregLike

[dmgreen]

Summary

This allows the peephole optimizer to know that a MVE_VMOV_to_lane_32 is the same as an insert subreg, allowing it to optimize some redundant lane moves.

[SjoerdMeijer]

llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp

5407

Hm, why this change? Is this a simplification? Don't think I understand this change.

[dmgreen]

llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp

5407

Thanks for taking a look.

VSETLNi32 handles D regs, so ssub_0 or ssub_1. Because MVE_VMOV_to_lane_32 is working on Q regs it can be ssub_2 or ssub_3 as well. There are other places in llvm that assume that ssub_0 + x == ssub_x, so that seems safe to use.

[SjoerdMeijer]

llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp

5407

Ah, okay, thanks, that's all clear now.

LGTM

llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp

 bool ARMBaseInstrInfo::getInsertSubregLikeInputs(
     const MachineInstr &MI, unsigned DefIdx, RegSubRegPair &BaseReg,
     RegSubRegPairAndIdx &InsertedReg) const {
   assert(DefIdx < MI.getDesc().getNumDefs() && "Invalid definition index");
   assert(MI.isInsertSubregLike() && "Invalid kind of instruction");
 
   switch (MI.getOpcode()) {
   case ARM::VSETLNi32:
+  case ARM::MVE_VMOV_to_lane_32:
     // dX = VSETLNi32 dY, rZ, imm
+    // qX = MVE_VMOV_to_lane_32 qY, rZ, imm
     const MachineOperand &MOBaseReg = MI.getOperand(1);
     const MachineOperand &MOInsertedReg = MI.getOperand(2);
     if (MOInsertedReg.isUndef())
       return false;
     const MachineOperand &MOIndex = MI.getOperand(3);
     BaseReg.Reg = MOBaseReg.getReg();
     BaseReg.SubReg = MOBaseReg.getSubReg();
 
     InsertedReg.Reg = MOInsertedReg.getReg();
     InsertedReg.SubReg = MOInsertedReg.getSubReg();
-    InsertedReg.SubIdx = MOIndex.getImm() == 0 ? ARM::ssub_0 : ARM::ssub_1;
+    InsertedReg.SubIdx = ARM::ssub_0 + MOIndex.getImm();
     return true;
   }
   llvm_unreachable("Target dependent opcode missing");
 }
 
 std::pair<unsigned, unsigned>
 ARMBaseInstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const {
   const unsigned Mask = ARMII::MO_OPTION_MASK;

llvm/lib/Target/ARM/ARMInstrMVE.td

class MVE_VMOV_lane_8<string suffix, bit U, MVE_VMOV_lane_direction dir>
   let Inst{22} = 0b1;
   let Inst{16} = Idx{3};
   let Inst{21} = Idx{2};
   let Inst{6} = Idx{1};
   let Inst{5} = Idx{0};
 }
 
 def MVE_VMOV_from_lane_32  : MVE_VMOV_lane_32<            MVE_VMOV_from_lane>;
-def MVE_VMOV_to_lane_32    : MVE_VMOV_lane_32<            MVE_VMOV_to_lane>;
 def MVE_VMOV_from_lane_s16 : MVE_VMOV_lane_16<"s16", 0b0, MVE_VMOV_from_lane>;
 def MVE_VMOV_from_lane_u16 : MVE_VMOV_lane_16<"u16", 0b1, MVE_VMOV_from_lane>;
-def MVE_VMOV_to_lane_16    : MVE_VMOV_lane_16< "16", 0b0, MVE_VMOV_to_lane>;
 def MVE_VMOV_from_lane_s8  : MVE_VMOV_lane_8 < "s8", 0b0, MVE_VMOV_from_lane>;
 def MVE_VMOV_from_lane_u8  : MVE_VMOV_lane_8 < "u8", 0b1, MVE_VMOV_from_lane>;
+let isInsertSubreg = 1 in
+def MVE_VMOV_to_lane_32    : MVE_VMOV_lane_32<            MVE_VMOV_to_lane>;
+def MVE_VMOV_to_lane_16    : MVE_VMOV_lane_16< "16", 0b0, MVE_VMOV_to_lane>;
 def MVE_VMOV_to_lane_8     : MVE_VMOV_lane_8 <  "8", 0b0, MVE_VMOV_to_lane>;
 
 // This is the same as insertelt but allows the inserted value to be an i32 as
 // will be used when it is the only legal type.
 def ARMVecInsert : SDTypeProfile<1, 3, [
   SDTCisVT<2, i32>, SDTCisSameAs<0, 1>, SDTCisPtrTy<3>
 ]>;
 def ARMinsertelt  : SDNode<"ISD::INSERT_VECTOR_ELT", ARMVecInsert>;

llvm/test/CodeGen/Thumb2/mve-vecreduce-add.ll

 }
 
 define arm_aapcs_vfpcc i64 @add_v2i16_v2i64_sext(<2 x i16> %x) {
 ; CHECK-LABEL: add_v2i16_v2i64_sext:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    vmov r0, s0
 ; CHECK-NEXT:    vmov r2, s2
 ; CHECK-NEXT:    sxth r0, r0
-; CHECK-NEXT:    vmov.32 q1[0], r0
 ; CHECK-NEXT:    asrs r1, r0, #31
-; CHECK-NEXT:    vmov.32 q1[1], r1
 ; CHECK-NEXT:    sxth r2, r2
-; CHECK-NEXT:    vmov r0, s4
 ; CHECK-NEXT:    adds r0, r0, r2
 ; CHECK-NEXT:    adc.w r1, r1, r2, asr #31
 ; CHECK-NEXT:    bx lr
 entry:
   %xx = sext <2 x i16> %x to <2 x i64>
   %z = call i64 @llvm.vector.reduce.add.v2i64(<2 x i64> %xx)
   ret i64 %z
 }

 }
 
 define arm_aapcs_vfpcc i64 @add_v2i8_v2i64_sext(<2 x i8> %x) {
 ; CHECK-LABEL: add_v2i8_v2i64_sext:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    vmov r0, s0
 ; CHECK-NEXT:    vmov r2, s2
 ; CHECK-NEXT:    sxtb r0, r0
-; CHECK-NEXT:    vmov.32 q1[0], r0
 ; CHECK-NEXT:    asrs r1, r0, #31
-; CHECK-NEXT:    vmov.32 q1[1], r1
 ; CHECK-NEXT:    sxtb r2, r2
-; CHECK-NEXT:    vmov r0, s4
 ; CHECK-NEXT:    adds r0, r0, r2
 ; CHECK-NEXT:    adc.w r1, r1, r2, asr #31
 ; CHECK-NEXT:    bx lr
 entry:
   %xx = sext <2 x i8> %x to <2 x i64>
   %z = call i64 @llvm.vector.reduce.add.v2i64(<2 x i64> %xx)
   ret i64 %z
 }

 }
 
 define arm_aapcs_vfpcc i64 @add_v2i16_v2i64_acc_sext(<2 x i16> %x, i64 %a) {
 ; CHECK-LABEL: add_v2i16_v2i64_acc_sext:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    vmov r2, s0
 ; CHECK-NEXT:    vmov r3, s2
 ; CHECK-NEXT:    sxth r2, r2
-; CHECK-NEXT:    vmov.32 q1[0], r2
-; CHECK-NEXT:    asrs r2, r2, #31
-; CHECK-NEXT:    vmov.32 q1[1], r2
+; CHECK-NEXT:    asr.w r12, r2, #31
 ; CHECK-NEXT:    sxth r3, r3
-; CHECK-NEXT:    vmov r12, s4
-; CHECK-NEXT:    adds.w r12, r12, r3
-; CHECK-NEXT:    adc.w r2, r2, r3, asr #31
-; CHECK-NEXT:    adds.w r0, r0, r12
-; CHECK-NEXT:    adcs r1, r2
+; CHECK-NEXT:    adds r2, r2, r3
+; CHECK-NEXT:    adc.w r3, r12, r3, asr #31
+; CHECK-NEXT:    adds r0, r0, r2
+; CHECK-NEXT:    adcs r1, r3
 ; CHECK-NEXT:    bx lr
 entry:
   %xx = sext <2 x i16> %x to <2 x i64>
   %z = call i64 @llvm.vector.reduce.add.v2i64(<2 x i64> %xx)
   %r = add i64 %z, %a
   ret i64 %r
 }
 

 }
 
 define arm_aapcs_vfpcc i64 @add_v2i8_v2i64_acc_sext(<2 x i8> %x, i64 %a) {
 ; CHECK-LABEL: add_v2i8_v2i64_acc_sext:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    vmov r2, s0
 ; CHECK-NEXT:    vmov r3, s2
 ; CHECK-NEXT:    sxtb r2, r2
-; CHECK-NEXT:    vmov.32 q1[0], r2
-; CHECK-NEXT:    asrs r2, r2, #31
-; CHECK-NEXT:    vmov.32 q1[1], r2
+; CHECK-NEXT:    asr.w r12, r2, #31
 ; CHECK-NEXT:    sxtb r3, r3
-; CHECK-NEXT:    vmov r12, s4
-; CHECK-NEXT:    adds.w r12, r12, r3
-; CHECK-NEXT:    adc.w r2, r2, r3, asr #31
-; CHECK-NEXT:    adds.w r0, r0, r12
-; CHECK-NEXT:    adcs r1, r2
+; CHECK-NEXT:    adds r2, r2, r3
+; CHECK-NEXT:    adc.w r3, r12, r3, asr #31
+; CHECK-NEXT:    adds r0, r0, r2
+; CHECK-NEXT:    adcs r1, r3
 ; CHECK-NEXT:    bx lr
 entry:
   %xx = sext <2 x i8> %x to <2 x i64>
   %z = call i64 @llvm.vector.reduce.add.v2i64(<2 x i64> %xx)
   %r = add i64 %z, %a
   ret i64 %r
 }
 

llvm/test/CodeGen/Thumb2/mve-vecreduce-mla.ll

 
 define arm_aapcs_vfpcc i64 @add_v2i64_v2i64(<2 x i64> %x, <2 x i64> %y) {
 ; CHECK-LABEL: add_v2i64_v2i64:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    .save {r4, lr}
 ; CHECK-NEXT:    push {r4, lr}
 ; CHECK-NEXT:    vmov r0, s4
 ; CHECK-NEXT:    vmov r1, s0
-; CHECK-NEXT:    vmov r2, s5
+; CHECK-NEXT:    vmov r3, s5
 ; CHECK-NEXT:    vmov r4, s7
-; CHECK-NEXT:    umull r12, r3, r1, r0
-; CHECK-NEXT:    mla r1, r1, r2, r3
+; CHECK-NEXT:    umull r12, r2, r1, r0
+; CHECK-NEXT:    mla r1, r1, r3, r2
 ; CHECK-NEXT:    vmov r2, s1
-; CHECK-NEXT:    vmov r3, s2
-; CHECK-NEXT:    vmov.32 q2[0], r12
+; CHECK-NEXT:    mla lr, r2, r0, r1
+; CHECK-NEXT:    vmov r0, s6
+; CHECK-NEXT:    vmov r2, s2
+; CHECK-NEXT:    umull r3, r1, r2, r0
+; CHECK-NEXT:    mla r1, r2, r4, r1
+; CHECK-NEXT:    vmov r2, s3
 ; CHECK-NEXT:    mla r1, r2, r0, r1
-; CHECK-NEXT:    vmov r2, s6
-; CHECK-NEXT:    vmov.32 q2[1], r1
-; CHECK-NEXT:    vmov r12, s8
-; CHECK-NEXT:    umull lr, r0, r3, r2
-; CHECK-NEXT:    mla r0, r3, r4, r0
-; CHECK-NEXT:    vmov r3, s3
-; CHECK-NEXT:    mla r2, r3, r2, r0
-; CHECK-NEXT:    adds.w r0, r12, lr
-; CHECK-NEXT:    adcs r1, r2
+; CHECK-NEXT:    adds.w r0, r12, r3
+; CHECK-NEXT:    adc.w r1, r1, lr
 ; CHECK-NEXT:    pop {r4, pc}
 entry:
   %m = mul <2 x i64> %x, %y
   %z = call i64 @llvm.vector.reduce.add.v2i64(<2 x i64> %m)
   ret i64 %z
 }
 
 define arm_aapcs_vfpcc i32 @add_v4i32_v4i32_acc(<4 x i32> %x, <4 x i32> %y, i32 %a) {

 ; CHECK-LABEL: add_v2i64_v2i64_acc:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    .save {r4, r5, r6, lr}
 ; CHECK-NEXT:    push {r4, r5, r6, lr}
 ; CHECK-NEXT:    vmov r2, s4
 ; CHECK-NEXT:    vmov r3, s0
 ; CHECK-NEXT:    vmov r4, s5
 ; CHECK-NEXT:    vmov r6, s7
-; CHECK-NEXT:    umull r12, lr, r3, r2
-; CHECK-NEXT:    mla r3, r3, r4, lr
+; CHECK-NEXT:    umull lr, r12, r3, r2
+; CHECK-NEXT:    mla r3, r3, r4, r12
 ; CHECK-NEXT:    vmov r4, s1
-; CHECK-NEXT:    vmov.32 q2[0], r12
-; CHECK-NEXT:    mla r2, r4, r2, r3
-; CHECK-NEXT:    vmov r4, s6
-; CHECK-NEXT:    vmov r3, s2
-; CHECK-NEXT:    vmov.32 q2[1], r2
-; CHECK-NEXT:    vmov r12, s8
-; CHECK-NEXT:    umull lr, r5, r3, r4
-; CHECK-NEXT:    mla r3, r3, r6, r5
+; CHECK-NEXT:    mla r12, r4, r2, r3
+; CHECK-NEXT:    vmov r3, s6
+; CHECK-NEXT:    vmov r4, s2
+; CHECK-NEXT:    umull r2, r5, r4, r3
+; CHECK-NEXT:    mla r4, r4, r6, r5
 ; CHECK-NEXT:    vmov r5, s3
-; CHECK-NEXT:    adds.w r6, r12, lr
-; CHECK-NEXT:    mla r3, r5, r4, r3
-; CHECK-NEXT:    adcs r2, r3
-; CHECK-NEXT:    adds r0, r0, r6
-; CHECK-NEXT:    adcs r1, r2
+; CHECK-NEXT:    adds.w r2, r2, lr
+; CHECK-NEXT:    mla r3, r5, r3, r4
+; CHECK-NEXT:    adc.w r3, r3, r12
+; CHECK-NEXT:    adds r0, r0, r2
+; CHECK-NEXT:    adcs r1, r3
 ; CHECK-NEXT:    pop {r4, r5, r6, pc}
 entry:
   %m = mul <2 x i64> %x, %y
   %z = call i64 @llvm.vector.reduce.add.v2i64(<2 x i64> %m)
   %r = add i64 %z, %a
   ret i64 %r
 }
 

llvm/test/CodeGen/Thumb2/mve-vld3.ll

 ; f16
 
 define void @vld3_v2f16(<6 x half> *%src, <2 x half> *%dst) {
 ; CHECK-LABEL: vld3_v2f16:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    ldrd r2, r3, [r0]
 ; CHECK-NEXT:    ldr r0, [r0, #8]
 ; CHECK-NEXT:    vmov.32 q0[0], r2
+; CHECK-NEXT:    vmov.16 q2[0], r2
 ; CHECK-NEXT:    vmov.32 q0[1], r3
 ; CHECK-NEXT:    vmov.32 q0[2], r0
-; CHECK-NEXT:    vmov r0, s0
 ; CHECK-NEXT:    vmovx.f16 s12, s0
-; CHECK-NEXT:    vmov.16 q2[0], r0
 ; CHECK-NEXT:    vmovx.f16 s4, s1
 ; CHECK-NEXT:    vins.f16 s12, s2
 ; CHECK-NEXT:    vins.f16 s8, s4
-; CHECK-NEXT:    vmov r0, s1
 ; CHECK-NEXT:    vadd.f16 q1, q2, q3
-; CHECK-NEXT:    vmovx.f16 s8, s2
-; CHECK-NEXT:    vmov.16 q0[0], r0
-; CHECK-NEXT:    vins.f16 s0, s8
-; CHECK-NEXT:    vadd.f16 q0, q1, q0
+; CHECK-NEXT:    vmovx.f16 s0, s2
+; CHECK-NEXT:    vmov.16 q2[0], r3
+; CHECK-NEXT:    vins.f16 s8, s0
+; CHECK-NEXT:    vadd.f16 q0, q1, q2
 ; CHECK-NEXT:    vmov r0, s0
 ; CHECK-NEXT:    str r0, [r1]
 ; CHECK-NEXT:    bx lr
 entry:
   %l1 = load <6 x half>, <6 x half>* %src, align 4
   %s1 = shufflevector <6 x half> %l1, <6 x half> undef, <2 x i32> <i32 0, i32 3>
   %s2 = shufflevector <6 x half> %l1, <6 x half> undef, <2 x i32> <i32 1, i32 4>
   %s3 = shufflevector <6 x half> %l1, <6 x half> undef, <2 x i32> <i32 2, i32 5>

 ; CHECK-NEXT:    vmovx.f16 s12, s0
 ; CHECK-NEXT:    vmov.16 q1[2], r2
 ; CHECK-NEXT:    ldrd r2, r0, [r0, #16]
 ; CHECK-NEXT:    vins.f16 s12, s2
 ; CHECK-NEXT:    vmovx.f16 s0, s2
 ; CHECK-NEXT:    vmov.32 q2[0], r2
 ; CHECK-NEXT:    vmovx.f16 s13, s3
 ; CHECK-NEXT:    vmov.32 q2[1], r0
-; CHECK-NEXT:    vmov r2, s1
+; CHECK-NEXT:    vmov r0, s1
 ; CHECK-NEXT:    vmovx.f16 s16, s8
 ; CHECK-NEXT:    vins.f16 s13, s9
 ; CHECK-NEXT:    vins.f16 s5, s16
-; CHECK-NEXT:    vmov r0, s8
 ; CHECK-NEXT:    vadd.f16 q1, q1, q3
-; CHECK-NEXT:    vmov.16 q3[0], r2
+; CHECK-NEXT:    vmov.16 q3[0], r0
 ; CHECK-NEXT:    vins.f16 s12, s0
 ; CHECK-NEXT:    vmovx.f16 s0, s9
-; CHECK-NEXT:    vmov.16 q3[2], r0
+; CHECK-NEXT:    vmov.16 q3[2], r2
 ; CHECK-NEXT:    vins.f16 s13, s0
 ; CHECK-NEXT:    vadd.f16 q0, q1, q3
 ; CHECK-NEXT:    vmov r2, s1
 ; CHECK-NEXT:    vmov r0, s0
 ; CHECK-NEXT:    strd r0, r2, [r1]
 ; CHECK-NEXT:    vpop {d8}
 ; CHECK-NEXT:    bx lr
 entry:

llvm/test/CodeGen/Thumb2/mve-vst3.ll

 ; CHECK-NEXT:    ldrd r4, r0, [r0, #16]
 ; CHECK-NEXT:    vmov q1[2], q1[0], lr, r3
 ; CHECK-NEXT:    vmov.32 q0[0], r4
 ; CHECK-NEXT:    vmov q1[3], q1[1], r12, r2
 ; CHECK-NEXT:    vmov.32 q0[1], r0
 ; CHECK-NEXT:    vmov.f32 s8, s7
 ; CHECK-NEXT:    vmov.f32 s10, s1
 ; CHECK-NEXT:    vmov r2, s8
-; CHECK-NEXT:    vmov r0, s10
 ; CHECK-NEXT:    vmov.f64 d4, d2
 ; CHECK-NEXT:    vmov.f32 s9, s6
 ; CHECK-NEXT:    vmov.f32 s10, s0
 ; CHECK-NEXT:    vmov.f32 s11, s5
 ; CHECK-NEXT:    vstrw.32 q2, [r1]
 ; CHECK-NEXT:    strd r2, r0, [r1, #16]
 ; CHECK-NEXT:    pop {r4, pc}
 entry:

 
 define void @vst3_v2f16(<2 x half> *%src, <6 x half> *%dst) {
 ; CHECK-LABEL: vst3_v2f16:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    vldmia r0, {s8, s9}
 ; CHECK-NEXT:    ldr r2, [r0, #8]
 ; CHECK-NEXT:    vmovx.f16 s12, s8
 ; CHECK-NEXT:    vins.f16 s8, s9
-; CHECK-NEXT:    vmov.32 q0[0], r2
 ; CHECK-NEXT:    vmov q1, q2
-; CHECK-NEXT:    vmov r2, s0
 ; CHECK-NEXT:    vmovx.f16 s8, s9
 ; CHECK-NEXT:    vmov.16 q1[2], r2
+; CHECK-NEXT:    vmov.32 q0[0], r2
 ; CHECK-NEXT:    vmov r0, s8
 ; CHECK-NEXT:    vins.f16 s5, s12
-; CHECK-NEXT:    vmovx.f16 s0, s0
 ; CHECK-NEXT:    vmov.16 q1[4], r0
+; CHECK-NEXT:    vmovx.f16 s0, s0
 ; CHECK-NEXT:    vins.f16 s6, s0
 ; CHECK-NEXT:    vmov r0, s4
 ; CHECK-NEXT:    vmov r2, s5
 ; CHECK-NEXT:    vmov r3, s6
 ; CHECK-NEXT:    stm r1!, {r0, r2, r3}
 ; CHECK-NEXT:    bx lr
 entry:
   %s1 = getelementptr <2 x half>, <2 x half>* %src, i32 0

 ; CHECK-NEXT:    vmov.32 q1[1], r12
 ; CHECK-NEXT:    vmov.32 q2[1], r2
 ; CHECK-NEXT:    vmov.32 q0[0], r4
 ; CHECK-NEXT:    vmov.f32 s6, s8
 ; CHECK-NEXT:    vmov.32 q0[1], r0
 ; CHECK-NEXT:    vmov.f32 s7, s9
 ; CHECK-NEXT:    vmovx.f16 s12, s4
 ; CHECK-NEXT:    vins.f16 s4, s6
-; CHECK-NEXT:    vmov r0, s0
-; CHECK-NEXT:    vmov q2, q1
 ; CHECK-NEXT:    vmovx.f16 s14, s6
-; CHECK-NEXT:    vmov.16 q2[2], r0
-; CHECK-NEXT:    vmov r0, s14
+; CHECK-NEXT:    vmov q2, q1
+; CHECK-NEXT:    vmov r2, s14
+; CHECK-NEXT:    vmov.16 q2[2], r4
 ; CHECK-NEXT:    vins.f16 s9, s12
 ; CHECK-NEXT:    vmovx.f16 s12, s0
-; CHECK-NEXT:    vmov.16 q2[4], r0
-; CHECK-NEXT:    vmov r2, s1
+; CHECK-NEXT:    vmov.16 q2[4], r2
+; CHECK-NEXT:    vmovx.f16 s0, s1
 ; CHECK-NEXT:    vins.f16 s10, s12
 ; CHECK-NEXT:    vmovx.f16 s12, s5
 ; CHECK-NEXT:    vins.f16 s5, s7
-; CHECK-NEXT:    vmovx.f16 s0, s1
 ; CHECK-NEXT:    vmovx.f16 s4, s7
 ; CHECK-NEXT:    vmov.f32 s11, s5
-; CHECK-NEXT:    vmov r0, s4
-; CHECK-NEXT:    vmov.16 q1[0], r2
+; CHECK-NEXT:    vmov r2, s4
+; CHECK-NEXT:    vmov.16 q1[0], r0
 ; CHECK-NEXT:    vins.f16 s4, s12
 ; CHECK-NEXT:    vstrw.32 q2, [r1]
-; CHECK-NEXT:    vmov.16 q1[2], r0
+; CHECK-NEXT:    vmov.16 q1[2], r2
 ; CHECK-NEXT:    vins.f16 s5, s0
 ; CHECK-NEXT:    vmov r2, s4
 ; CHECK-NEXT:    vmov r0, s5
 ; CHECK-NEXT:    strd r2, r0, [r1, #16]
 ; CHECK-NEXT:    pop {r4, pc}
 entry:
   %s1 = getelementptr <4 x half>, <4 x half>* %src, i32 0
   %l1 = load <4 x half>, <4 x half>* %s1, align 4